ChatGPT

The name overpromises and underpromises at the same time. This node doesn't open a ChatGPT-style window and it won't chat with you directly - but it is the piece that lets your whole NodeGPT workflow think like GPT. It wraps the OpenAI API into something AutoGen can talk to, and hands it to the agents downstream. Think of it as the "LLM config provider" for the cloud model option, sitting next to the local ones (LM Studio, Ollama, llama-cpp, oobabooga).

NodeGPT is an early experiment that pulled Microsoft's AutoGen multi-agent framework into ComfyUI's node graph. Every agent node in the pack - Assistant, Engineer, Critic, Planner, Executor - needs an LLM object on its input before it can say anything. This node produces exactly that object. You never see text come out of it; you see a configured brain that the agents then use.

What you actually set

Only two fields exist, and one of them you'll often leave alone:

• Model - defaults to gpt-4, which is both a sign of this pack's age (it launched in late 2023) and a prompt to change it. Set whatever OpenAI model name you have access to.
• API_Key - optional on the surface, mandatory in practice if you want it to work. Paste it in, or better, feed it from the LoadAPIconfig node.

The single output is LLM, and that's the whole game: wire it into any agent node, or straight into GroupChat if you're building the manager-based setup.

The API key problem, take it seriously

The README says it in all caps: "Dont share your workflow with any API keys inside!!!" That's not paranoia. ComfyUI embeds workflow JSON into output images by default, so a key typed into a widget travels into every PNG you save - and if you post that PNG, the key is right there in the metadata for anyone to grab. The pack even ships LoadAPIconfig so you can keep the key in a text file and pass it over a wire instead of hardcoding it into the graph. Use it.

Installing it

The README's path is the classic one:

cd ComfyUI/custom_nodes
git clone https://github.com/xXAdonesXx/NodeGPT

Restart ComfyUI and it should pull its requirements on startup (the pack depends on Microsoft's pyautogen at its core). One honest caveat: as of this writing the original repo returns 404 - the author's GitHub account is gone - so that clone URL is dead and ComfyUI Manager can't install it either. The surviving fork (Guizmus/NodeGPT) only carries the pack's early text-gen build, not these AutoGen nodes. If you load a NodeGPT workflow today you'll likely be chasing missing-node errors. None of that changes what this node taught the ecosystem: how an LLM provider, an agent, and a text output get wired into a diffusion graph.

Troubleshooting

• Nothing happens / empty output - the key is wrong, the model name isn't one you're allowed to call, or there's no network. Check the terminal; errors land there, not in the UI.
• Chat runs forever - this is a known pack-wide quirk; AutoGen's default termination logic wasn't wired up cleanly, so a chat can spin until you kill the queue.
• Keep expectations calibrated. The author said it plainly: "not everything is functioning correctly yet."